Dataset for Certifying adaptive experimentation from decision-time information

This dataset provides the complete author-generated evidence and manuscript source data accompanying the study “Certifying adaptive experimentation from decision-time information.” The study introduces Opportunity-aware Policy Authorization for Laboratories (OPAL), a framework for authorizing adaptive experimental capacity only when a precommitted contract jointly supports opportunity, finite-population certifiability, risk control, executed value, non-trivial activation and cost. The deposit preserves the distinct evidential roles of retrospective analysis, exact theory, held-out simulation, offline measured-outcome evaluation and physical-archive constructibility. The record contains: source data and publication-ready files for all Main and Extended Data figures; the main manuscript, Supplementary Information, bibliography and figure-generation programs; frozen simulation ledgers and exact finite-population certification outputs; held-out false-activation calibration and validation results; retrospective comparator and execution-cost analyses; derived Causal Chambers constructibility evidence; frozen CTRP partitions, assignments and locked exact-outcome evaluation tables; cpg0012 Cell Painting design manifests, frozen target-calibrated model, sealed pre-outcome assignments, final outcomes, primary metrics and common-contract comparator battery; integrity reports, SHA-256 manifests and a result-to-source map linking each manuscript claim to its authoritative evidence file. In the single-held-out cpg0012 final evaluation, the frozen policy activated 595 of 11,265 compounds, including 384 positive opportunities. Six of seven deliberately stringent internal 95% checks passed. The false-activation upper confidence bound was 5.18%, the positive-sensitivity lower confidence bound was 5.41%, and the executed-value BCa lower confidence bound was strictly positive at 0.001229. The false-discovery point estimate was 34.62%, below the internal 35% target, but its one-sided 95% upper confidence bound was 37.97%; consequently, the complete joint certificate was not awarded. A read-only independent audit reproduced the final result and passed all 46 integrity and numerical checks. The CTRP and cpg0012 studies are offline experimental-capacity abstractions, not clinical treatment-selection studies or prospective physical-instrument deployments. The Causal Chambers analysis establishes computational constructibility only and does not demonstrate causal physical-policy value. Raw CTRP, cpg0012 Cell Painting and Causal Chambers archives are not redistributed. The deposit instead provides pinned source identifiers, provenance information, checksums, author-generated partitions, derived result ledgers and all result-bearing materials required to verify the reported statistics. Reproducible analysis code is provided separately in the associated GitHub and Zenodo software records. Version 1.1.0 adds the target-calibrated cpg0012 final-test evidence, the sealed 11,265-unit assignment record, final comparator results, the 46-check audit, revised manuscript source data and the updated Main Figure 6.